Low overhead virtual machines tracing in a cloud infrastructure

Transcription

1 Low overhead virtual machines tracing in a cloud infrastructure Mohamad Gebai Michel Dagenais Dec 7, 2012 École Polytechnique de Montreal

2 Content Area of research Current tracing: LTTng vs ftrace / virtio trace Preliminary work Main project proposal References 2

3 Area of research Tracing virtualized machines in a cloud context 3

4 Current VM tracing LTTng tools 2.1: support for network streaming Sending trace data between the guest and the host over the network Might not be the best suited approach for the guest and the host to communicate trace data Usage of network bandwidth IP based communications may be inadvertently blocked by the virtual machine administrator 4

5 Virtio virtio/figure2.gif Based on paravirtualization : the operating system knows it is running in a VM Virtio: a set of device drivers for virtual machines virtio net (network adapter) virtio balloon (dynamic allocation and disallocation of the guest's memory) virtio serial (char device communication channel) 5

6 Virtio trace Device driver for tracing since Linux kernel version 3.7 Sending traces from guest to host without copying Low overhead Guest OS tracing tool Trace data No memory copying Ring buffer read( ) virtio trace splice( ) virtio serial FIFO qemu Host OS 6

7 Preliminary work Running benchmarks on native VM, with ftrace (using virtio trace), with LTTng (using streaming) Compare the guest to host communication overhead between ftrace and LTTng 7

8 Main project proposal Improve the performance and extent of virtual machine tracing Adapt the best features of existing kernel ring buffers for LTTng and ftrace Insure uniform clock sources between kernel and user space applications in physical and virtual machines Trace and analyze VM specific behavior such as page sharing with KSM (kernel same page merging) 8

9 References virtio/index.html serial/

10 Large Scale Data Center Monitoring and Debugging Infrastructure Julien Desfossez Michel Dagenais Dec 7, 2012 École Polytechnique de Montreal

11 Content Problem definition State of the art Proposed approach Early work References 11

12 Problem Definition Large scale data centers : hosting, cloud computing, scientific computing, etc Virtualized data centers Monitoring inefficient and costly (/proc reading, locks and system calls) Mostly high level metrics (CPU, memory, load, etc) or application level monitoring without correlation to the OS or physical layer No virtualisation/hypervisor metrics 12

13 Monitoring Multiple aspects: Availability testing Load measurement Trending Reporting 13

14 State of the Art Commercial products add Cloud to their softwares titles Most of commercial systems focus on the interface and a predefined set of tests Reuse the same technology as monitoring a single machine and the associated services Focus mainly on capacity planning with high level information 14

15 State of the Art Swift Recon : middle man metrics collector Common metrics : Load average, sockets stats, /proc/meminfo, etc Object storage dedicated metrics : MD5 of each ring file most recent object replication time count of each type of quarantined files: account, container, or object. Count of async_pendings (deferred container updates) on disk 15

16 State of the Art Rackspace Cloud Monitoring API to access monitoring data High level remote tests HTTP requests to pages TCP connect Banner check 16

17 State of the Art RockSteady Metric analysis and correlation engine User defined metrics Sending/receiving done with message queueing (RabbitMQ) Optional graphing engine (Graphite) Complex Event Processing to extract root cause from event by integrating multiple data sources 17

18 18

19 Proposed Approach Tracing can be used as an efficient monitoring backend (Master's results) LTTng + Perf PMU counters Extract metrics from trace data Define metrics relevant for large scale data centers and virtualized environments Generate statistics and detailled usage reports 19

20 Proposed Approach Dynamically adjustable level of information (high level metrics down to detailled in application behavior) Streaming and live analysis Aggregation on intermediate nodes (depending on topology) 20

21 Target environment OpenStack KVM LXC cgroups 21

22 Early Work Basic metrics computation and reporting (LTTngTop during Master) Streaming traces (lttng tools 2.1) Index generation Live analysis (in progress) 22

23 References OpenStack Manual chapter 6, OpenStack Object Storage Monitoring Proven Practice: Metrics for Virtualization Management ?decorator Intel Virtualization Performance Metrics us/articles/virtualization performa Google Rocksteady 23